California has created a state designation regime for third-party AI auditors and told courts that an audit performed to a standard identified under the new chapter is relevant to, but not conclusive of, an action alleging that a defendant's development, modification or use of an AI system caused harm

What the chapter sets up

Section 8898 supplies the definitions. An AI auditor is a person, partnership, academic institution, nonprofit or corporation that conducts a covered audit on behalf of a third party. A covered AI audit is one conducted to assess internal controls, processes or systems implemented for an AI system or model that are necessary for compliance with state law. An independent verification organization, or IVO, is an AI auditor designated by the Government Operations Agency as having demonstrated expertise in assessing the risks posed by an AI system or model and identifying the metrics and methodologies that form the basis for that assessment.

Section 8898.1 puts eight deliverables on the agency, all of them due on or before 1 January 2028. Application requirements come first, and an applicant's submission covers its qualifications, information addressing the published designation criteria, the benchmarks, technologies, metrics and methodologies it proposes to use, and whatever documentation the agency needs to verify the accuracy of the application.

Suspension and termination procedures are next, and the chapter names six things those procedures are to consider: failures to adhere to appropriate standards, material misrepresentations in the application, audit reports or required disclosures, conflicts of interest that impair independence, failure to maintain adequate documentation, conduct that reasonably calls into question the integrity, objectivity or competence of the IVO, and lapses in cybersecurity.

The designation criteria at subdivision (c) carry the substance. One sentence there deserves quoting in full, because it is the closest the chapter comes to a hard prohibition: an IVO "may accept payment from a party being assessed at reasonable market rates but shall not accept terms which condition any payment or the amount of any payment on the results of their assessment". Independence is treated separately, and includes having no operational or management dependence on the assessed party or its affiliates and remaining otherwise free from that party's control in reaching conclusions or making recommendations.

The rest of Section 8898.1 is housekeeping with teeth in the tail: publish requirements and criteria in a publicly accessible format, publish a prominent statement that publishing them is not a state endorsement of any AI system or model, review and revise them as law and practice move, align with existing professional and regulatory audit and assurance standards to the extent practicable, and structure the requirements to minimise duplicative compliance obligations, including by letting work done for substantially similar requirements count here.

The evidentiary sentence, read with the four disclaimers above it

Section 8898.4(a) lists four things the chapter does not do. It does not establish liability solely for failure to comply with a standard under the chapter. It does not constitute a state recommendation or endorsement of any AI system or model. It does not require anyone developing, deploying or operating an AI system or model to engage an IVO or undergo a covered AI audit as a condition of doing any of those things in California. And it does not require an IVO to conduct audits assessing compliance with applicable state law in order to register with the agency.

Then subdivision (b): "In an action alleging that a defendant's development, modification, or use of an artificial intelligence system or model caused harm, the fact that an audit has been performed in accordance with a standard identified under this chapter is relevant to, but not conclusive of, the action."

Read what that sentence does not contain. It creates no presumption. It shifts no burden. It does not define harm, does not name a cause of action, and does not limit itself to tort. It does not say which party may introduce the fact of the audit, and relevance is not one-directional: in an action where an audit was performed and the harm happened anyway, the same sentence makes that fact admissible too.

Calling this a safe harbour would misstate it in both directions. Subdivision (a)(1) stops a plaintiff from grounding liability solely on a failure to meet a chapter standard, and subdivision (b) stops a defendant from treating a conforming audit as dispositive. What the Legislature has enacted is an evidentiary hook, with the weight left to the trier of fact.

Our reading, marked as ours: the practical effect arrives later than the chaptered date suggests. No standard is identified under the chapter until the agency does the work Section 8898.1 assigns it, and that work is due on or before 1 January 2028.

Annual reporting, redaction, and a five-year retention duty

Section 8898.3 puts an annual report on each designated IVO, filed with both the agency and the Legislature, and not sooner than 12 months after initial designation. It covers summaries of the IVO's standards and methodologies, a description of any changes to its governance policies or funding sources relevant to conflicts of interest or independence, and any changes to its application information.

Redaction is permitted where necessary to protect trade secrets, cybersecurity, public safety or United States national security, or to comply with federal or state law. Two conditions ride with it. The IVO describes the character and justification of the redaction in any published version of the document, to the extent the concerns justifying the redaction permit. And it retains the unredacted information for five years.

That retention period is the part an assurance firm should cost. Five years of unredacted working material, held against a possible future request, is an archiving obligation rather than a filing obligation.

Working groups, and who has to be in the room

Section 8898.2(a) lists nine categories of stakeholder the agency consults as appropriate, running from AI auditors and academic institutions through private entities that develop AI and private entities that deploy or operate it, to consumer protection, labour and civil society organisations, other agencies and standards-setting bodies.

Subdivision (b) is firmer than the consultation list. The agency convenes working groups to solicit stakeholder input in identifying standards and developing and revising the procedures, requirements and criteria, and those groups are to include, at a minimum, engineers from AI companies that are competitors and AI safety experts. Putting competing firms' engineers in the same statutory room is a deliberate design choice, and it is the kind of provision that generates its own antitrust-adjacent questions in practice. The chapter says nothing about how those are to be handled.

The agency then reports to the Legislature on the working groups' findings under Section 9795.

Three cross-references that do not point where they say

Section 8898.1(e) requires a published statement that "the publication of requirements or criteria pursuant to this subdivision" is not a state endorsement. The publication duty sits in subdivision (d), not in (e).

Section 8898.3(c) opens "If an IVO redacts information in a document pursuant to this subdivision". The redaction permission sits in subdivision (b).

Section 8898.4(a)(4) refers to what is needed "in order to register with the agency", while every other provision in the chapter speaks of designation, application and designation criteria. The chapter sets up no register.

None of this is fatal, and courts read past drafting slips of this kind routinely. We record them because anyone drafting a compliance memo off the chaptered text will hit all three, and because the fourth item we noticed is one we are not asserting at all. See the note below.

What we did not verify

What we opened: the chaptered text of SB 813 on the California Legislative Counsel site, the version stamped Chaptered 09/09/26 and published 09/09/2026 09:00 PM, read end to end, including the Legislative Counsel's Digest, the enactment clause, and Sections 8898 through 8898.4.

What we did not open: the bill PDF, the eleven earlier versions, the committee analyses, the vote record, the Transparency in Frontier Artificial Intelligence Act that the Digest describes as existing law, Government Code Section 9795 which governs the report to the Legislature, and the Department of Technology inventory provisions the Digest also describes. We describe existing law only as the Digest describes it, and the Digest is the Legislative Counsel's summary rather than operative text.

One structural observation we are not asserting. In the rendering we read, the application requirements in Section 8898.1(a) run straight to subparagraphs (A) through (D) with no intervening paragraph level. That may be an artefact of how the page renders rather than the enrolled structure, and we did not open the PDF to check, so anyone citing that subdivision should confirm the citation form against the official copy.

What we refuse to claim: we do not say a conforming audit provides a defence, because Section 8898.4(b) makes it relevant but not conclusive and says nothing about burden. We do not say the chapter regulates AI developers, because it regulates an agency and the auditors it designates. We give no commencement date for the chapter, because the chaptered text carries no commencement or urgency clause and we did not examine California's general rule on when statutes take effect. We name no standard as identified under the chapter, because none exists until the agency acts. Quotations are reproduced with ASCII punctuation in place of the typographic apostrophes used in the published text, which is a house typesetting convention and not a change to any word.

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.

Related briefings

• California's SB 53 frontier AI transparency law in force
• Illinois SB 315 and frontier AI audit duties
• NYC Local Law 144 bias audit enforcement and employer risk
• North America AI regulation tracker
• AI Regulation News hub